learning by near-miss an example of using

1
learning by near-missan example of using
codingknowledge
2
preamble...

• learns concept models
• real objects/events/etc coded as Kn
• (following example uses tuples)
• model is refined using examples
• ve examples relax/generalise
• -ve examples restrict/specialise

3
1. initial example

• (isa x block)
• (isa y block)
• (isa z block)
• (supports y x)
• (supports z x)
• (pos x horis)
• (pos y vert )
• (pos z vert )

4
2. -ve example

• difference
• (supports y x)
• (supports z x)
• changes
• (imp supports y x)
• (imp supports z x)

5
3. another -ve example

• differences
• (pos x horis)
• (not touches y z)
• changes
• (imp pos x horis)
• note use of...
• 'general Kn
• most important diff.s

6
4. another -ve example

• differences
• (not touches y z)
• changes
• (imp not touches y z)

7
5. a ve example

• differences
• (isa x wedge)
• changes
• (isa x (wedge block))

8
the refined description

• refined
• (isa x (wedge block))
• (isa y block)
• (isa z block)
• (imp supports y x)
• (imp supports z x)
• (imp pos x horis)
• (pos y vert )
• (pos z vert )
• (imp not touches y z)

• original
• (isa x block)
• (isa y block)
• (isa z block)
• (supports y x)
• (supports z x)
• (pos x horis)
• (pos y vert )
• (pos z vert )

9
the process

• compare new old descriptions
• if ve example generalise
• express diffs in terms of new
• select most sig diffs
• extend old by diff list
• else if ve example specialise
• express diffs in terms of old
• select most sig diffs
• enforce old by diff list

10
comparing representations
11
comparing representations

• simple approaches try all matches
• better approaches best 1st search

12
using best 1st search

• start with open labelling
• add new label with each successor state
• rank diffs to generate diff score
• explore state with min diff score